WhiteLine Daily: Chip stocks enter a bear market, and Apple briefly reclaimed the top spot in market value thanks to lighter AI investment

WhiteLine Daily is a new column launched by Wu Shuo Blockchain, focusing on the global markets and technology industry in the AI era. It tracks how these changes affect the crypto market and asset pricing, and distills the most important daily market clues in a way that’s easier to understand.

01｜Macro Snapshot

The chip index enters a bear market, and AI trading begins to deleverage

On Friday, U.S. equities continued to adjust lower. The S&P 500 fell 1.0%, the Nasdaq fell 1.4%, and the Dow Jones fell 0.8%. Driven by the Middle East situation, Brent crude rose 4.6%, though U.S. Treasury yields pulled back somewhat.

The most obvious sign in this round of adjustment is still in chip stocks. The Philadelphia Semiconductor Index fell about 10% over the week, retreating more than 20% from the June highs—technically entering a bear market. A triple-long semiconductor ETF is down more than half from its peak, suggesting leveraged capital is facing greater selling pressure.

But for the year, the chip index is still up cumulatively more than 60%. The first squeeze from this pullback is the group of stocks that have risen fastest and have the most crowded positioning—so we still can’t directly conclude that AI demand has already weakened.

WhiteLine’s assessment:

Chip stocks could see an oversold rebound at any time, but a decline of more than 20% alone can’t confirm a bottom. Next, we need to watch whether leveraged ETFs stop mechanically selling, whether institutions have finished trimming positions, and whether cloud providers continue to maintain their capital expenditure guidance.

02｜Valuation Rotation

Apple briefly returns to the No. 1 spot by market cap, and the market starts rewarding a lighter AI path

On July 17, Apple’s market cap briefly reached about $4.88 trillion, surpassing Nvidia’s roughly $4.86 trillion, temporarily reclaiming the top spot globally by market value. This ranking will move with the stock price, but it reflects loosening market preferences.

Over the past year, the market has most favored chips, data center investment, and compute capacity spending. Apple has kept its investment relatively restrained in these areas, but it has the iPhone, service subscriptions, and a massive user entry point. As long as AI features can drive handset upgrades or increase service revenue, it has a chance to deliver returns with a lower level of investment.

WhiteLine’s assessment:

The market previously mainly rewarded companies that provide chips and compute power, but now it’s starting to place greater emphasis on user entry points, monetization capability, and cash flow. In the next stage, companies with controllable capital expenditures and clear monetization paths are more likely to receive a valuation premium.

03｜Model Efficiency

Kimi K3 makes the market re-calculate: how much compute is needed for the same level of AI capability

Moonshot AI’s Kimi K3 has 2.8 trillion parameters and a 1 million token context window. Although the model is large, when processing each token it only activates 16 of the 896 experts. Kimi officially states that this architecture improves overall expansion efficiency by about 2.5 times compared with Kimi K2.

The official API price for Kimi K3 is: $0.3 per million tokens for cached input, $3 for regular input, and $15 for output. Full model weights are planned to be released on July 27. The performance results published so far mainly come from tests by Kimi itself; the full technical report and third-party evaluations still need to be awaited. Therefore, at this stage, what can be confirmed is the model size, architecture, and pricing—while real-world usage experience still requires more external validation.

WhiteLine’s assessment:

Kimi K3’s impact is not only about parameter scale—it once again drives down the cost of using frontier models. Cheaper models will bring more applications, and they will also keep pressure on model companies’ pricing room. In the future, the more practical benchmark is: how much does it really cost to complete a real task.

04｜AI Infrastructure

Compute begins to be rented out, and data center expansion also meets more resistance

Citing people familiar with the matter, Reuters reported that Meta is considering renting out compute capacity to Anthropic, with a potential deal worth up to $10 billion over two years. However, negotiations are still at an early stage; neither side has officially confirmed, and an agreement may ultimately not be reached. Separately, the Wall Street Journal reported that SpaceX is discussing a compute collaboration deal worth tens of billions of dollars with the U.S. Department of Defense. SpaceX and the Pentagon did not respond, and Reuters said it could not independently verify it—so this remains a negotiation-only item.

What’s more noteworthy are parts that have already materialized: SpaceX and Google have signed a multi-year cloud services agreement involving about 110,000 Nvidia chips; Anthropic has also confirmed that it will use compute capacity from SpaceX Colossus 1 data centers totaling about 300 megawatts.

Meanwhile, data center expansion is facing community pushback. Event organizers plan to hold protests in at least 125 locations in the United States. A Reuters and Ipsos survey shows that only 14% of respondents are willing to support building AI data centers in their own communities.

WhiteLine’s assessment:

The valuation focus for AI data centers is shifting from the number of GPUs to device utilization rates, power supply, and construction permits. Infrastructure owners that secure long-term customers and power resources will be more valuable. Expansion plans lacking orders or approval support may gradually become a valuation burden.

Today’s Main Theme

All four pieces of information today point to the same change: how much revenue AI investment can ultimately generate.

Falling chip stocks are cleaning up excessive leverage; Apple is back in focus thanks to lighter investment and direct, user-facing charging/monetization entry points; Kimi K3 continues to push down model prices; and Meta and SpaceX are trying to turn compute into a service that can be sold to the outside.

Going forward, the market will care more about several specific things: the cost of completing tasks, data center utilization rates, the long-term contracts that have already been signed, and the electricity and approval requirements that can truly be implemented.